Re: [Jfs-discussion] [Bug 9291] pdflush stuck in D state with v2.6.24-rc1-192-gef49c32

[Fengguang Wu <wf...@ma...>]

On Mon, Dec 24, 2007 at 03:13:56PM +0800, Fengguang Wu wrote:
> On Mon, Dec 24, 2007 at 10:25:53AM +0800, Fengguang Wu wrote:
> > On Sun, Dec 23, 2007 at 10:35:45AM -0800, bug...@bu... wrote:
> > > http://bugzilla.kernel.org/show_bug.cgi?id=9291
> > 
> > Hmm, I just tried JFS on LVM - still OK.
> > It seems not related to LVM.
> 
> I can now reproduce the bug on JFS with the following command:
> 
> debootstrap --arch i386 etch /mnt/jfs http://debian.ustc.edu.cn/debian
> 
> It's a rather compound procedure, but I just cannot trigger the bug through
> simple operations like cp/concatenate/truncate ...
> 
> The symptoms:
> 
> - one pdflush stuck in D state:
> 
> USER       PID %CPU %MEM    VSZ   RSS TTY      STAT START   TIME COMMAND
> root       321  0.0  0.0      0     0 ?        D    13:45   0:01 [pdflush]
> root     15397  0.0  0.0      0     0 ?        S    14:21   0:00 [pdflush]

It was confirmed that the source of this bug lies in metapage_writepage():

                if (!mp || !test_bit(META_dirty, &mp->flag))
                        continue;

That logic skips the following line:

                        set_page_writeback(page);

which should be called to clear the PAGECACHE_TAG_DIRTY tag.

The META_dirty bit could be cleared in several places, e.g.
__invalidate_metapages().

Any ideas about a solution?